shi_xx

Results 11 issues of shi_xx

any suggestions for optimization? 1. Speed up the reasoning 2. improve the accuracy of segmentation details thx!

报错信息: [ERROR] Exported quantized model can not be loaded, only deployment is supported. 来源应该是这里吧: https://github.com/PaddlePaddle/PaddleX/blob/07ceb75823fd9d4eb1aa9e5fd93ab636e801507c/paddlex/cv/models/load_model.py#L117

https://github.com/deepinsight/insightface/blob/e78eee59caff3515a4db467976cf6293aba55035/parsing/dml_csr/dataset/datasets.py#L179 return 的结果缺少 semantic_edges? https://github.com/deepinsight/insightface/blob/e78eee59caff3515a4db467976cf6293aba55035/parsing/dml_csr/train.py#L302 这里最后的 _ 应该是meta 吧 缺少了semantic_edges,可以描述下怎么生成的吗? 谢谢

https://github.com/chenjun2hao/DDRNet.pytorch/blob/bc0e193e87ead839dbc715c48e6bfb059cf21b27/lib/utils/utils.py#L35 "and" should be "or"

## Change _Choose one_ - Bugfix - New feature - Chore - Tests - Refactoring ## Description of changes _Please describe your changes_ ## Screenshots _Please include screenshots for new...

看res2net的layer1部分:26*4 = 104 ==conv1x1==> 256 想减小通道数,把256 改成 64 请问,可以把26改成13,或者更小吗?会有什么不好的影响吗 比如下面: ① 13*4 = 52 ==conv1x1==> 64 ② 7*4 = 28 ==conv1x1==> 64 谢谢

Detail 模块中关注的是边缘细节,线条居多 dice loss 对 area 效果更好些,用在 detail 模块是否合适? 目测 dice loss 的波动比较大,改成其他 boundary aware loss 是否会更好?

I am unable to change the place by dbclick??

您好,用 AGRNet 在 celebahqmask 上跑了几十个epoch后,遇到些问题,请教下: 1. pred_edge 的效果跟 label 比起来,相差太多(参考图片) 2. 测试图片时,一般虚化背景会比较好,如果背景干扰一多,对解析的结果影响蛮大 3. 小目标(比如人眼等)结果有点糟糕 请问,是训练问题吗,还是模型本身的限制?有什么优化办法吗? 谢谢! ![图片](https://user-images.githubusercontent.com/5567783/177471519-a349b7fa-087f-4b5e-8735-9ec6c0418f17.png)

rt. 少了几个大写字母,VXYZ